def get_model(data_dir,
mat_prop,
classification=False,
batch_size=None,
transfer=None,
verbose=True):
# Get the TorchedCrabNet architecture loaded
model = Model(CrabNet(compute_device=compute_device).to(compute_device),
model_name=f'{mat_prop}',
verbose=verbose)
# Train network starting at pretrained weights
if transfer is not None:
model.load_network(f'{transfer}.pth')
model.model_name = f'{mat_prop}'
# Apply BCEWithLogitsLoss to model output if binary classification is True
if classification:
model.classification = True
# Get the datafiles you will learn from
train_data = f'{data_dir}/{mat_prop}/train.csv'
try:
val_data = f'{data_dir}/{mat_prop}/val.csv'
except:
print('Please ensure you have train (train.csv) and validation data',
f'(val.csv) in folder "data/materials_data/{mat_prop}"')
# Load the train and validation data before fitting the network
data_size = pd.read_csv(train_data).shape[0]
batch_size = 2**round(np.log2(data_size) - 4)
if batch_size < 2**7:
batch_size = 2**7
if batch_size > 2**12:
batch_size = 2**12
model.load_data(train_data, batch_size=batch_size, train=True)
print(f'training with batchsize {model.batch_size} '
f'(2**{np.log2(model.batch_size):0.3f})')
model.load_data(val_data, batch_size=batch_size)
# Set the number of epochs, decide if you want a loss curve to be plotted
model.fit(epochs=40, losscurve=False)
# Save the network (saved as f"{model_name}.pth")
model.save_network()
return model
def get_model(mat_prop,
i,
classification=False,
batch_size=None,
transfer=None,
verbose=True):
# Get the TorchedCrabNet architecture loaded
model = Model(CrabNet(compute_device=compute_device).to(compute_device),
model_name=f'{mat_prop}{i}',
verbose=verbose)
# Train network starting at pretrained weights
if transfer is not None:
model.load_network(f'{transfer}.pth')
model.model_name = f'{mat_prop}'
# Apply BCEWithLogitsLoss to model output if binary classification is True
if classification:
model.classification = True
# Get the datafiles you will learn from
train_data = rf'data\matbench_cv\{mat_prop}\train{i}.csv'
val_data = rf'data\matbench_cv\{mat_prop}\val{i}.csv'
# Load the train and validation data before fitting the network
data_size = pd.read_csv(train_data).shape[0]
batch_size = 2**round(np.log2(data_size) - 4)
if batch_size < 2**7:
batch_size = 2**7
if batch_size > 2**12:
batch_size = 2**12
# batch_size = 2**7
model.load_data(train_data, batch_size=batch_size, train=True)
print(f'training with batchsize {model.batch_size} '
f'(2**{np.log2(model.batch_size):0.3f})')
model.load_data(val_data, batch_size=batch_size)
# Set the number of epochs, decide if you want a loss curve to be plotted
model.fit(epochs=300, losscurve=False)
# Save the network (saved as f"{model_name}.pth")
model.save_network()
return model
